A model is never the brain. In Allternit the system is the agent. Models are interchangeable cognitive components that the system calls for specific steps. State, authority, lifecycle and completion belong to the system. If swapping one model for another changes the agent loop, the state model, tool authority, completion semantics or the workflow graph, the implementation does not conform to the architecture.
Architecture contract
What an agent is
An agent is one external identity backed by a system:- a typed task and agent state record, not a transcript;
- explicit, scoped, expiring authority checked by policy;
- a durable run carried by the work runtime;
- completion decided from verification evidence;
- receipts for every effect, input and verdict.
Terms
The product metaphor
Allternit’s product story is “One Brain, Many Faces, One Nervous System”. It still describes the product, and it sits under the rule above:- The Brain is the whole runtime system (the kernel and the Gizzi runtime), not a model.
- The Faces are the surfaces: desktop, web, extensions, CLI. They hold no agent logic of their own.
- The Nervous System is the SDK and communication layer that connects surfaces to the runtime.